Navigation

Account
Loading account...
The Cognitive Toolkit

Your Thinking Has Failure Modes. These Tools Catch Them.

Muddy thinking. Skipped steps. Missing perspectives. Unweighted trade-offs. Each tool catches a different failure mode before you ship it.

The Tools

Each tool is designed for a specific failure mode. Use them when your main AI workflow is stuck, when you need to challenge your assumptions, or when decisions need documented reasoning.

Structured Reflection

The bug was obvious — once you said it out loud.

You've been staring at it for an hour. Explain it to something that asks questions instead of giving answers. What you actually say is rarely what you thought you were thinking. Sessions persist for when it resurfaces.

  • Persistent sessions
  • Multiple styles
  • Convergence detection

Chain of Thought

You jumped straight to the solution. The problem was wrong.

Stage gates that won't let you skip steps. Define the problem before researching. Research before analyzing. Catch where you leapt to conclusions before the code is written.

  • Staged progression
  • Context preservation
  • Progress tracking

Context Switcher

The flaw was obvious — from every perspective but yours.

Multiple stakeholder perspectives in parallel. Technical, business, user, risk, ops — plus templates for architecture, security review, debugging, API design, and feature evaluation. Surface blind spots you can’t see from your own viewpoint.

  • Parallel analysis
  • Customizable views
  • Synthesis engine

Decision Matrix

Your gut had a favorite. Your AI validated it in seconds.

Weighted criteria. Scored options. Sensitivity analysis. See if your decision survives changing priorities — not gut feelings validated by a helpful assistant.

  • Weighted criteria
  • Parallel evaluation
  • Scoring with reasoning

Devil's Advocate

Everyone agreed. That's exactly when you're most exposed.

Structured adversarial analysis. Pre-mortems, assumption challenges, and attack trees that hunt for the failure you’re not looking at — before it finds you in production.

  • Pre-mortem analysis
  • Assumption challenges
  • Attack-tree mapping

Graph of Thought

One path forward is a guess. Explore, score, then merge the best.

Diverge into multiple approaches, score each on its merits, and synthesize the strongest into something better than any single path. Linear thinking picks one and hopes; this branches, evaluates, and converges.

  • Divergent generation
  • Scored evaluation
  • Aggregation

Formal Logic

Not "that sounds right." Proven, or disproven.

Deterministic verification for your reasoning. Entailment, satisfiability, and temporal-logic checks return a verdict with a witness — not a confidence score. This is where the model stops guessing and the math answers.

  • Entailment & proof
  • Satisfiability
  • Temporal (LTL) checks

Hindsight

"I knew it all along." No, you didn’t — and that’s the lesson.

Post-mortems that resist hindsight bias. Compare what you actually predicted against what happened, and extract lessons that are calibrated — not outcome knowledge dressed up as foresight.

  • Prediction vs. outcome
  • Bias detection
  • Calibrated lessons

Not Prompts. Tools.

Most "reasoning" wrappers inject a system prompt the model can ignore. We inject structured tools that create auditable reasoning trails.

1

Isolated Sessions

Each session runs in isolation. Your main context stays clean—you get the insight, nothing else.

2

Observable Tool Calls

The LLM invokes cognitive tools: score_option,surface_contradiction. You see every call.

3

Persistent Sessions

Sessions are saved automatically. Mark key moments. Return to past insights. Build a library of your reasoning—not just the conclusions.

# Decision Matrix session - the LLM gets these tools:
tools = [
    define_criteria,      # Lock in what matters before evaluating
    set_weights,          # Make priorities explicit
    score_option,         # Evaluate with reasoning attached
    run_sensitivity,      # Check if your decision is robust
]

# Sample tool invocation (observable in session):
> INVOKE: score_option(
    option="PostgreSQL",
    criterion="consistency_guarantees",
    score=0.95,
    reasoning="Full ACID compliance, strong transaction support"
)

Your AI Won't Challenge You. These Will.

$20/month for reasoning sessions that challenge instead of validate. No usage metering, no overage charges.